---
id: con-jagged-frontier
schema: concept/1
layer: garden
status: established
name: Jagged frontier
aka: ["uneven AI abilities", "where AI is good and bad", "capabilities and limits", "spiky capability"]
summary: >
  The uneven line between tasks AI does well and tasks it does badly, which often doesn't match
  human intuition about what is hard, so you have to test it on your own tasks.
first_seen: "2023-09"
relations:
  builds_on: [con-large-language-model]
  evolved_from: []
  related: [con-cognitive-offloading, con-human-in-the-loop]
  parallels:
    - idea: "a spiky skills profile"
      field: "education"
      note: "Some learners are far ahead in one area and behind in a neighbouring one; averages hide it."
    - idea: "knowing a colleague's strengths before delegating"
      field: "management"
implementations: []
competencies: [cmp-check-the-result, cmp-frame-the-outcome]
origin:
  kind: ai-proposed
  by: "seed agent 2026-10-16"
  rationale: "Explains why trust must be task-specific; drawn from a widely cited 2023 field experiment with consultants."
---

AI's abilities form a jagged edge. It may write a strong strategy memo and then miscount the items in a list. It may translate a contract clause well and invent a detail about your town. Tasks that feel similar to us can fall on opposite sides of the line.

The phrase comes from a 2023 study with management consultants. On tasks inside AI's abilities, consultants using AI were faster and better. On a task just outside them, consultants using AI were more likely to get it wrong than those working alone, and most didn't notice.

The edge also moves as models improve, so last year's limits may not apply. So map the frontier for your own work: try real tasks, notice where it's reliable and where it slips, and keep checking in the places it slipped.

## Example

A recruiter finds AI excellent at drafting job descriptions and summarising interview notes, but unreliable at counting years of experience from CVs. She keeps using it for the first two and does the counting herself.

## How it connects

- Explains why **model choice** and **verification** must be task by task.
- **Hallucination** is hardest to catch on tasks you haven't tested yet.
- Deciding what to hand over is a **cognitive offloading** choice.

## Try it

List five tasks you might use AI for and mark each "AI will nail it" or "AI will struggle". Try two from each group and see which guesses were wrong.